LibWasm: Implement the rest of the SIMD conversions

Author: https://github.com/dzfrias
Commit: 4b9649282e
Pull-request: https://github.com/LadybirdBrowser/ladybird/pull/719
Reviewed-by: https://github.com/alimpfard

Userland/Libraries/LibWasm/AbstractMachine/BytecodeInterpreter.cpp

@@ -1701,9 +1701,9 @@ void BytecodeInterpreter::interpret(Configuration& configuration, InstructionPoi
     case Instructions::v128_store64_lane.value():
         return pop_and_store_lane_n<64>(configuration, instruction);
     case Instructions::i32x4_trunc_sat_f32x4_s.value():
-        return unary_operation<u128, u128, Operators::VectorFloatConvertOp<4, Operators::SaturatingTruncate<i32>>>(configuration);
+        return unary_operation<u128, u128, Operators::VectorConvertOp<4, 4, u32, f32, Operators::SaturatingTruncate<i32>>>(configuration);
     case Instructions::i32x4_trunc_sat_f32x4_u.value():
-        return unary_operation<u128, u128, Operators::VectorFloatConvertOp<4, Operators::SaturatingTruncate<u32>>>(configuration);
+        return unary_operation<u128, u128, Operators::VectorConvertOp<4, 4, u32, f32, Operators::SaturatingTruncate<u32>>>(configuration);
     case Instructions::i8x16_bitmask.value():
         return unary_operation<u128, i32, Operators::VectorBitmask<16>>(configuration);
     case Instructions::i16x8_bitmask.value():
@@ -1725,18 +1725,21 @@ void BytecodeInterpreter::interpret(Configuration& configuration, InstructionPoi
     case Instructions::i16x8_q15mulr_sat_s.value():
         return binary_numeric_operation<u128, u128, Operators::VectorIntegerBinaryOp<8, Operators::SaturatingOp<i16, Operators::Q15Mul>, MakeSigned>>(configuration);
     case Instructions::f32x4_convert_i32x4_s.value():
-        return unary_operation<u128, u128, Operators::VectorIntegerConvertOp<4, Operators::Convert<f32>, MakeSigned>>(configuration);
+        return unary_operation<u128, u128, Operators::VectorConvertOp<4, 4, u32, i32, Operators::Convert<f32>>>(configuration);
     case Instructions::f32x4_convert_i32x4_u.value():
-        return unary_operation<u128, u128, Operators::VectorIntegerConvertOp<4, Operators::Convert<f32>, MakeUnsigned>>(configuration);
+        return unary_operation<u128, u128, Operators::VectorConvertOp<4, 4, u32, u32, Operators::Convert<f32>>>(configuration);
-    case Instructions::f32x4_demote_f64x2_zero.value():
-    case Instructions::f64x2_promote_low_f32x4.value():
-    case Instructions::i32x4_trunc_sat_f64x2_s_zero.value():
-    case Instructions::i32x4_trunc_sat_f64x2_u_zero.value():
     case Instructions::f64x2_convert_low_i32x4_s.value():
+        return unary_operation<u128, u128, Operators::VectorConvertOp<2, 4, u64, i32, Operators::Convert<f64>>>(configuration);
     case Instructions::f64x2_convert_low_i32x4_u.value():
-        dbgln_if(WASM_TRACE_DEBUG, "Instruction '{}' not implemented", instruction_name(instruction.opcode()));
+        return unary_operation<u128, u128, Operators::VectorConvertOp<2, 4, u64, u32, Operators::Convert<f64>>>(configuration);
-        m_trap = Trap { ByteString::formatted("Unimplemented instruction {}", instruction_name(instruction.opcode())) };
+    case Instructions::f32x4_demote_f64x2_zero.value():
-        return;
+        return unary_operation<u128, u128, Operators::VectorConvertOp<4, 2, u32, f64, Operators::Convert<f32>>>(configuration);
+    case Instructions::f64x2_promote_low_f32x4.value():
+        return unary_operation<u128, u128, Operators::VectorConvertOp<2, 4, u64, f32, Operators::Convert<f64>>>(configuration);
+    case Instructions::i32x4_trunc_sat_f64x2_s_zero.value():
+        return unary_operation<u128, u128, Operators::VectorConvertOp<4, 2, u32, f64, Operators::SaturatingTruncate<i32>>>(configuration);
+    case Instructions::i32x4_trunc_sat_f64x2_u_zero.value():
+        return unary_operation<u128, u128, Operators::VectorConvertOp<4, 2, u32, f64, Operators::SaturatingTruncate<u32>>>(configuration);
     }
 }
 

Userland/Libraries/LibWasm/AbstractMachine/Operators.h

@@ -882,51 +882,33 @@ struct VectorFloatUnaryOp {
     }
 };
 
-template<size_t VectorSize, typename Op>
+template<size_t ResultSize, size_t InputSize, typename ResultType, typename InputType, typename Op>
-struct VectorFloatConvertOp {
+struct VectorConvertOp {
     auto operator()(u128 lhs) const
     {
-        using VectorInput = NativeFloatingVectorType<128, VectorSize, NativeFloatingType<128 / VectorSize>>;
+        using VectorInput = NativeVectorType<128 / InputSize, InputSize, MakeUnsigned>;
-        using VectorResult = NativeVectorType<128 / VectorSize, VectorSize, MakeUnsigned>;
+        using VectorResult = NativeVectorType<128 / ResultSize, ResultSize, MakeUnsigned>;
         auto value = bit_cast<VectorInput>(lhs);
         VectorResult result;
         Op op;
-        for (size_t i = 0; i < VectorSize; ++i) {
+        auto size = min(InputSize, ResultSize);
-            result[i] = op(value[i]);
+        for (size_t i = 0; i < size; ++i)
+            result[i] = bit_cast<ResultType>(op(bit_cast<InputType>(value[i])));
+        // FIXME: We shouldn't need this, but the auto-vectorizer sometimes doesn't see that we
+        // need to pad with zeroes when InputSize < ResultSize (i.e. converting from f64x2 -> f32x4).
+        // So we put this here to make sure. Putting [[clang::optnone]] over this function resolves
+        // this issue, but that would be pretty unacceptable...
+        if constexpr (InputSize < ResultSize) {
+            constexpr size_t remaining = ResultSize - InputSize;
+            for (size_t i = 0; i < remaining; ++i)
+                result[i + InputSize] = 0;
         }
         return bit_cast<u128>(result);
     }
 
     static StringView name()
     {
-        switch (VectorSize) {
+        switch (ResultSize) {
-        case 4:
-            return "vecf(32x4).cvt_op"sv;
-        case 2:
-            return "vecf(64x2).cvt_op"sv;
-        default:
-            VERIFY_NOT_REACHED();
-        }
-    }
-};
-
-template<size_t VectorSize, typename Op, template<typename> typename SetSign = MakeSigned>
-struct VectorIntegerConvertOp {
-    auto operator()(u128 lhs) const
-    {
-        using VectorInput = NativeVectorType<128 / VectorSize, VectorSize, SetSign>;
-        using VectorResult = NativeFloatingVectorType<128, VectorSize, NativeFloatingType<128 / VectorSize>>;
-        auto value = bit_cast<VectorInput>(lhs);
-        VectorResult result;
-        Op op;
-        for (size_t i = 0; i < VectorSize; ++i)
-            result[i] = op(value[i]);
-        return bit_cast<u128>(result);
-    }
-
-    static StringView name()
-    {
-        switch (VectorSize) {
         case 4:
             return "vec(32x4).cvt_op"sv;
         case 2: